Text-to-Image Synthesis for Architectural Design: Generating Floor Plans using BERT

Architectural design is a challenging field with limited AI applicability due to the difficulty of translating textual descriptions into visual representations. This work fills the gap by proposing a multi-modal learning framework that combines textual and visual data for AI-based architectural automation. Using BERT embeddings for text and numerical arrays for spatial composition, the approach entails data preprocessing, normalization, and structured alignment to bridge the AI-architectural intent gap. A Python-based pipeline controls text-image relationships, making it usable for automated design generation, classification, and retrieval. Experimental validation using SSIM, and expert assessment, confirms the framework's capability to generate functionally consistent and visually accurate layouts. The results show an AI-assisted design process that improves accuracy, streamlines conceptualization, and enables greater AI-architecture integration. Future work will improve attention mechanisms, increase datasets, and investigate more comprehensive architectural applications.